A deep neural network approach for learning intrinsic protein-RNA binding preferences

Area of Science:

• Computational biology
• Bioinformatics
• Molecular biology

Background:

• Protein-RNA complexes are crucial for numerous biological processes, including gene regulation and viral replication.
• Accurate prediction of protein-RNA interactions is essential for understanding cellular dynamics and disease mechanisms.
• Computational approaches are needed to efficiently infer protein-RNA binding models for novel interactions.

Purpose of the Study:

• To develop a novel deep learning approach, DLPRB, for predicting protein-RNA binding.
• To investigate the utility of joint analysis of RNA sequence and structure for binding prediction.
• To compare the performance of convolutional neural networks (CNNs) and recurrent neural networks (RNNs) for this task.

Main Methods:

• Developed DLPRB, a deep neural network (DNN) framework utilizing both CNN and RNN architectures.
• Incorporated RNA sequence and structural context (represented as probability vectors) as input features.
• Trained and evaluated models on high-throughput in vitro and in vivo protein-RNA binding data.

Main Results:

• DLPRB demonstrated substantial improvements in predicting in vitro protein-RNA binding compared to existing methods.
• A statistically significant, albeit more modest, improvement was observed for in vivo binding prediction.
• Utilizing experimentally-measured RNA structure further enhanced prediction accuracy for in vivo data.

Conclusions:

• DLPRB offers a powerful new tool for accurate protein-RNA binding prediction.
• Jointly analyzing RNA sequence and structure significantly improves binding prediction models.
• The developed method provides biological insights into protein-RNA binding mechanisms.